Fairness Guardrails

1) Definition

• Fairness guardrails = constraints or checks that ensure a model’s predictions are equitable across subgroups (e.g., gender, race, age, geography).
• They prevent a model from being accurate overall but unfair to specific populations.

Example: A loan approval model might have 90% accuracy, but only 60% recall for women vs 90% for men → unfair, even if global AUC looks good.

---

2) Why fairness guardrails matter

• Ethical reasons: avoid discrimination.
• Legal reasons: compliance with regulations (e.g., EEOC, GDPR, Equal Credit Opportunity Act).
• Business reasons: protect reputation, ensure trust, reduce risk of biased outcomes.

---

3) Common Fairness Metrics

a) Group fairness metrics (population-level)

• Demographic parity: prediction rates should be similar across groups.
  • $P(\hat{Y}=1 \mid A=a) = P(\hat{Y}=1 \mid A=b)$
• Equal opportunity: true positive rates (recall) equal across groups.
• Equalized odds: both TPR and FPR equal across groups.
• Calibration within groups: predicted probabilities mean the same thing for each subgroup.

b) Individual fairness metrics

• Similar individuals should get similar predictions.
• Harder to enforce, but important in high-stakes settings.

---

4) Example Fairness Guardrails

• Recall parity guardrail:
  • Gap in recall between subgroups ≤ 3 percentage points.
• Approval rate guardrail:
  • Predicted positive rate (loans approved) for women must not be less than 95% of that for men.
• Calibration guardrail:
  • For each subgroup, predicted 70% probability → observed success rate should be ~70%.

---

5) Implementation workflow

1. Define sensitive attributes (gender, race, age, location, etc.).
2. Choose fairness metrics relevant to your application.
3. Set thresholds (guardrails). Example: recall gap ≤ 3pp.
4. Evaluate on validation/test sets.
5. Monitor in production (data drift can cause fairness to degrade over time).
6. Take action if violated → retrain, reweight, or adjust decision threshold.

---

6) Techniques to improve fairness

• Pre-processing: balance dataset (reweight, resample).
• In-processing: fairness-constrained optimization (regularization, adversarial debiasing).
• Post-processing: adjust thresholds or outputs per subgroup to meet fairness guardrails.

---

Summary

• Fairness guardrails = rules to ensure equitable performance across groups.
• Key metrics: demographic parity, equal opportunity, equalized odds, calibration.
• Guardrails are expressed as thresholds on fairness metrics (e.g., ≤ 3pp subgroup gap).
• Used in hiring, lending, healthcare, and any sensitive ML deployment.